OpenBrainLM — a biomimetic brain harness with hippocampal memory for LLMs

[0SxD]

Relevant to Khoj

Khoj is an AI second brain. OpenBrainLM is a brain harness that structures HOW the brain thinks and remembers. Different layers, potentially complementary.

What OpenBrainLM does

Long-term memory modeled on the human hippocampus:

• Three-tier memory: working memory → short-term buffer → long-term storage
• Consolidation cycle (like biological sleep) — replays session activity, promotes verified findings, strengthens connections that worked (Hebbian learning)
• Quarantine layer (blood-brain barrier) — new information is quarantined until verified by an immune agent. Nothing enters long-term memory unverified.
• Brain regions — knowledge organized by domain, routed by semantic salience (not keyword search). The routing improves with use.
• Append-only — knowledge is never deleted, like a ledger.

Trinity dialectic for reasoning:

• Three voices debate every significant decision: Ethos (evidence corridor + criteria), Logos (evaluation), Pathos (mission)
• Based on Aristotle's Nicomachean Ethics — structured disagreement as the reasoning mechanism
• 9 sub-evaluators (Trinity of Trinities) catch blind spots the main debate misses

Brainstem for discipline:

• Hooks fire before/after every action
• Inhibition-by-default (basal ganglia model) — actions are released, not activated
• Research corridors with verification gates

Self-learning:

• NEAT neuroevolution evolves the agent routing topology during sleep
• The brain that wakes up has better wiring than the brain that went to sleep

Every component maps to real biology with citations — octopus arm autonomy, hippocampal sharp-wave ripples, fruit fly neural circuits, basal ganglia inhibition.

LLM-agnostic, pure Python stdlib, MIT licensed.

Repo: https://github.com/0SxD/OpenBrainLM

The memory architecture is detailed in concept paper 2. Would love feedback from the Khoj community — you're thinking about the same problems.